youmuliu/stgsharp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

no progress in matrix parallel

Browse Source
This commit is contained in:
Youmu Liu
2025-09-01 22:17:52 +08:00
parent 5e68a1e786
commit 2475a3b0ed
8 changed files with 224 additions and 904 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Common/HighPerformance/Memory/SlabAllocator.cs
+21 -19
View File
@@ -59,25 +59,27 @@ namespace StgSharp.HighPerformance.Memory
/// Thrown when requesting unsupported configurations. /// Thrown when requesting unsupported configurations.
/// </exception> /// </exception>
/// <remarks> /// <remarks>
/// <para><strong> Important Threading Considerations: </strong></para> <para> While SLAB /// <para><strong> Important Threading Considerations: </strong></para> <para>While SLAB
/// allocators themselves may be designed for single-threaded use, allocated objects are /// /// allocators themselves may be designed for single-threaded use, allocated objects
/// frequently shared across multiple threads in real-world applications. Due to the /// are /// frequently shared across multiple threads in real-world applications. Due
/// special implementation of expandable concurrent sequential SLAB allocators, they rely /// to the /// special implementation of expandable concurrent sequential SLAB
/// on internal expansion locks to ensure that memory buffer expansions (which involve /// allocators, they rely /// on internal expansion locks to ensure that memory buffer
/// copying the entire buffer to a new location) do not occur while other threads are /// expansions (which involve /// copying the entire buffer to a new location) do not
/// reading from allocated regions.</para> <para><strong> Sequential Layout Limitations: /// occur while other threads are /// reading from allocated regions. </para>
/// </strong><br /> Sequential layout SLAB allocators cannot provide a single-threaded /// <para><strong>Sequential Layout Limitations: ///</strong><br /> Sequential layout SLAB
/// version because:<list type="bullet"><item><description> They use contiguous memory /// allocators cannot provide a single-threaded /// version because:<list
/// layout that requires buffer expansion when capacity is exceeded /// type="bullet"><item><description>They use contiguous memory /// layout that requires
/// </description></item><item><description> Buffer expansion involves copying all /// buffer expansion when capacity is exceeded
/// existing data to a new memory location </description></item><item><description> /// ///</description></item><item><description>Buffer expansion involves copying all ///
/// Without proper synchronization, concurrent access to allocated objects during /// existing data to a new memory location</description></item><item><description> ///
/// expansion would cause memory corruption </description></item><item><description> The /// Without proper synchronization, concurrent access to allocated objects during ///
/// expansion lock mechanism is essential even if allocation itself is single-threaded /// expansion would cause memory corruption</description></item><item><description>The ///
/// </description></item></list></para> <para><strong> Chunked Layout: </strong><br /> /// expansion lock mechanism is essential even if allocation itself is single-threaded
/// Chunked layout allocators can optionally provide single-threaded versions since they /// ///</description></item></list></para> <para><strong> Chunked Layout: </strong><br />
/// don't require buffer expansion and copying. Each chunk is allocated independently, /// /// Chunked layout allocators can optionally provide single-threaded versions since
/// avoiding the memory relocation issues present in sequential layouts.</para> /// they /// don't require buffer expansion and copying. Each chunk is allocated
/// independently, /// avoiding the memory relocation issues present in sequential
/// layouts. </para>
/// </remarks> /// </remarks>
public static SlabAllocator<T> Create(nuint count, SlabBufferLayout layout, bool concurrentSupport = true) public static SlabAllocator<T> Create(nuint count, SlabBufferLayout layout, bool concurrentSupport = true)
{ {
StgSharp.Common/Mathematics/Linear/Matrix.Basic.cs
+2
View File
@@ -46,6 +46,8 @@ namespace StgSharp.Mathematics
throw new IndexOutOfRangeException("Matrix dimensions must agree."); throw new IndexOutOfRangeException("Matrix dimensions must agree.");
} }
Matrix<float> result = GCMatrix<float>.FromDefault(left.ColumnLength, left.RowLength); Matrix<float> result = GCMatrix<float>.FromDefault(left.ColumnLength, left.RowLength);
MatrixParallel.PublishTask();
return result;
} }
} }
StgSharp.Common/Mathematics/Linear/Matrix.cs
+69 -69
View File
@@ -123,94 +123,94 @@ namespace StgSharp.Mathematics
} }
} }
public readonly unsafe ref struct MatrixSegmentEnumeration( }
MatrixKernel* source,
int stride, public readonly unsafe ref struct MatrixSegmentEnumeration(
int size, MatrixKernel* source,
int secondaryStride, int stride,
int secondarySize) int size,
int secondaryStride,
int secondarySize)
{
private readonly MatrixKernel* _source = source;
private readonly int _secondarySize = secondarySize;
private readonly int _secondaryStride = secondaryStride;
private readonly int _size = size;
private readonly int _stride = stride;
public Enumerator GetEnumerator()
{
return new Enumerator(in this);
}
public ref struct Enumerator(in MatrixSegmentEnumeration enumeration)
{ {
private readonly MatrixKernel* _source = source; private MatrixKernel* _current = enumeration._source - enumeration._stride;
private readonly int _secondarySize = secondarySize; private readonly MatrixKernel* _end = enumeration._source + enumeration._size;
private readonly int _secondaryStride = secondaryStride; private readonly int _secondarySize = enumeration._secondarySize;
private readonly int _secondaryStride = enumeration._secondaryStride;
private readonly int _stride = enumeration._stride;
private readonly int _size = size; public readonly Segment Current
private readonly int _stride = stride;
public Enumerator GetEnumerator()
{ {
return new Enumerator(in this); [MethodImpl(MethodImplOptions.AggressiveInlining)]
get => new(_current, _secondarySize, _secondaryStride);
} }
public ref struct Enumerator(in MatrixSegmentEnumeration enumeration) public bool MoveNext()
{ {
_current += _stride;
private MatrixKernel* _current = enumeration._source - enumeration._stride; return _current < _end;
private readonly MatrixKernel* _end = enumeration._source + enumeration._size;
private readonly int _secondarySize = enumeration._secondarySize;
private readonly int _secondaryStride = enumeration._secondaryStride;
private readonly int _stride = enumeration._stride;
public readonly Segment Current
{
[MethodImpl(MethodImplOptions.AggressiveInlining)]
get => new(_current, _secondarySize, _secondaryStride);
}
public bool MoveNext()
{
_current += _stride;
return _current < _end;
}
} }
} }
public readonly unsafe ref struct Segment(MatrixKernel* source, int size, int stride) }
public readonly unsafe ref struct Segment(MatrixKernel* source, int size, int stride)
{
private readonly MatrixKernel* _source = source;
private readonly int _size = size;
private readonly int _stride = stride;
public ref MatrixKernel this[int index]
{
get
{
if (index < 0 || index >= _size) {
throw new ArgumentOutOfRangeException(nameof(index));
}
return ref _source[index * _stride];
}
}
public Enumerator GetEnumerator()
{
return new Enumerator(_source, _size, _stride);
}
public ref struct Enumerator(MatrixKernel* source, int count, int stride)
{ {
private readonly MatrixKernel* _source = source; private readonly MatrixKernel* _end = source + (count * stride);
private readonly int _size = size;
private readonly int _stride = stride; private readonly int _stride = stride;
public ref MatrixKernel this[int index] public MatrixKernel* Current
{ {
get
{
if (index < 0 || index >= _size) {
throw new ArgumentOutOfRangeException(nameof(index));
}
return ref _source[index * _stride];
}
}
public Enumerator GetEnumerator()
{
return new Enumerator(_source, _size, _stride);
}
public ref struct Enumerator(MatrixKernel* source, int count, int stride)
{
private readonly MatrixKernel* _end = source + (count * stride);
private readonly int _stride = stride;
public MatrixKernel* Current
{
[MethodImpl(MethodImplOptions.AggressiveInlining)] [MethodImpl(MethodImplOptions.AggressiveInlining)]
get; get;
[MethodImpl(MethodImplOptions.AggressiveInlining)] [MethodImpl(MethodImplOptions.AggressiveInlining)]
private set; private set;
} = source - stride; } = source - stride;
public bool MoveNext()
{
Current += _stride;
return Current < _end;
}
public bool MoveNext()
{
Current += _stride;
return Current < _end;
} }
} }
StgSharp.Common/Mathematics/Linear/MatrixParallel.Task.cs
+15 -34
View File
@@ -35,60 +35,41 @@ using System.Drawing;
using System.Numerics; using System.Numerics;
using System.Runtime.CompilerServices; using System.Runtime.CompilerServices;
using System.Runtime.InteropServices; using System.Runtime.InteropServices;
using System.Threading;
namespace StgSharp.Mathematics namespace StgSharp.Mathematics
{ {
/// <summary> [StructLayout(LayoutKind.Explicit)]
/// Matrix parallel computation task structure with scalar data packet support Uses type- public unsafe struct MatrixParallelTaskWrap
/// erased design for thread pool compatibility
/// </summary>
[StructLayout(LayoutKind.Explicit, Size = 48)]
public unsafe struct MatrixParallelTask
{ {
public const int SizeofConstantParameter = sizeof(ulong);
public const int SizeOfTask = 48;
[FieldOffset(0)] public MatrixKernel* left; [FieldOffset(0)] public MatrixKernel* left;
[FieldOffset(8)] public MatrixKernel* result; [FieldOffset(8)] public MatrixKernel* result;
[FieldOffset(16)] public MatrixKernel* right; [FieldOffset(16)] public MatrixKernel* right;
[FieldOffset(24)] public int scalerCount; [FieldOffset(40)] public ScalarDataPacket* scalerPacket;
[FieldOffset(24)] public int ComputeMode;
[FieldOffset(60)] public int count;
[FieldOffset(48)] public int leftStride;
[FieldOffset(56)] public int resultStride;
[FieldOffset(52)] public int rightStride;
[FieldOffset(32)] public IntPtr executionHandle; [FieldOffset(32)] public IntPtr executionHandle;
/// <summary>
/// Pointer to the scalar data packet containing up to 8 parameters Each parameter is
/// stored as ulong (8 bytes) for type erasure
/// </summary>
[FieldOffset(40)] public nint scalerDataPacket;
/// <summary>
/// Execute the task using type-erased function pointer
/// </summary>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public readonly void Execute()
{
if (scalerCount == 0)
{
((delegate*<MatrixKernel*, MatrixKernel*, MatrixKernel*, void>)executionHandle)(left, right, result);
} else
{
((delegate*<MatrixKernel*, MatrixKernel*, MatrixKernel*, void*, void>)executionHandle)(left, right, result, (void*)scalerDataPacket);
}
}
} }
/// <summary> /// <summary>
/// Scalar data packet for storing up to 8 parameters as type-erased ulong values Used to /// Scalar data packet for storing up to 8 parameters as type-erased ulong values Supports
/// separate scalar data allocation from task allocation /// reference counting for shared usage across multiple tasks Used to separate scalar data
/// allocation from task allocation and track batch completion
/// </summary> /// </summary>
[StructLayout(LayoutKind.Explicit, Size = 72)]
public unsafe struct ScalarDataPacket public unsafe struct ScalarDataPacket
{ {
public const int MaxScalarCount = 8; public const int MaxScalarCount = 8;
public const int PacketSize = 64; public const int PacketSize = 64;
private fixed ulong Value[8]; [FieldOffset(0)] private fixed ulong Value[8];
public ScalarDataPacket() public ScalarDataPacket()
{ {
StgSharp.Common/Mathematics/Linear/MatrixParallel.TaskFactory.cs
-738
View File
@@ -1,738 +0,0 @@
//-----------------------------------------------------------------------
// -----------------------------------------------------------------------
// file="MatrixParallel.TaskFactory.cs"
// Project: StgSharp
// AuthorGroup: Nitload Space
// Copyright (c) Nitload Space. All rights reserved.
//
// Permission is hereby granted, free of charge, to any person
// obtaining a copy of this software and associated documentation
// files (the ¡°Software¡±), to deal in the Software without restriction,
// including without limitation the rights to use, copy, modify, merge,
// publish, distribute, sublicense, and/or sell copies of the Software,
// and to permit persons to whom the Software is furnished to do so,
// subject to the following conditions:
//
// The above copyright notice and
// this permission notice shall be included in all copies
// or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED ¡°AS IS¡±,
// WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED,
// INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
// IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM,
// DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
// ARISING FROM, OUT OF OR IN CONNECTION WITH
// THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
//
// -----------------------------------------------------------------------
// -----------------------------------------------------------------------
using StgSharp.HighPerformance;
using StgSharp.HighPerformance.Memory;
using System;
using System.Numerics;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
namespace StgSharp.Mathematics
{
/// <summary>
/// Factory for creating and managing MatrixParallelTask instances with pooled allocation Uses
/// separate SLAB allocators for tasks and scalar data packets
/// </summary>
public static unsafe class MatrixParallelTaskFactory
{
private static readonly SlabAllocator<ScalarDataPacket> ScalarPacketAllocator =
SlabAllocator<ScalarDataPacket>.Create(1024, SlabBufferLayout.Chunked, true);
// SLAB allocators for pooled allocation
private static readonly SlabAllocator<MatrixParallelTask> TaskAllocator =
SlabAllocator<MatrixParallelTask>.Create(2048, SlabBufferLayout.Chunked, true);
/// <summary>
/// Create a task with no scalar parameters
/// </summary>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static MatrixTaskHandle CreateTask(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle)
{
nuint taskPtr = TaskAllocator.Allocate();
MatrixParallelTask* task = (MatrixParallelTask*)taskPtr;
task->left = left;
task->right = right;
task->result = result;
task->executionHandle = executionHandle;
task->scalerCount = 0;
task->scalerDataPacket = 0; // null pointer
return new MatrixTaskHandle(task, null, false);
}
/// <summary>
/// Create a task with 1 scalar parameter
/// </summary>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static MatrixTaskHandle CreateTask<T0>(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
T0 scaler0)
where T0: unmanaged, INumber<T0>
{
return CreateTaskInternal(left, right, result, executionHandle, 1, &InitializePacket<T0>, scaler0);
}
/// <summary>
/// Create a task with 2 scalar parameters
/// </summary>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static MatrixTaskHandle CreateTask<T0, T1>(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
T0 scaler0,
T1 scaler1)
where T0: unmanaged, INumber<T0>
where T1: unmanaged, INumber<T1>
{
return CreateTaskInternal(
left, right, result, executionHandle, 2, &InitializePacket<T0, T1>, scaler0, scaler1);
}
/// <summary>
/// Create a task with 3 scalar parameters
/// </summary>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static MatrixTaskHandle CreateTask<T0, T1, T2>(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
T0 scaler0,
T1 scaler1,
T2 scaler2)
where T0: unmanaged, INumber<T0>
where T1: unmanaged, INumber<T1>
where T2: unmanaged, INumber<T2>
{
return CreateTaskInternal(
left, right, result, executionHandle, 3, &InitializePacket<T0, T1, T2>, scaler0, scaler1, scaler2);
}
/// <summary>
/// Create a task with 4 scalar parameters
/// </summary>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static MatrixTaskHandle CreateTask<T0, T1, T2, T3>(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
T0 scaler0,
T1 scaler1,
T2 scaler2,
T3 scaler3)
where T0: unmanaged, INumber<T0>
where T1: unmanaged, INumber<T1>
where T2: unmanaged, INumber<T2>
where T3: unmanaged, INumber<T3>
{
return CreateTaskInternal(
left, right, result, executionHandle, 4, &InitializePacket<T0, T1, T2, T3>, scaler0, scaler1, scaler2,
scaler3);
}
/// <summary>
/// Create a task with 5 scalar parameters
/// </summary>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static MatrixTaskHandle CreateTask<T0, T1, T2, T3, T4>(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
T0 scaler0,
T1 scaler1,
T2 scaler2,
T3 scaler3,
T4 scaler4)
where T0: unmanaged, INumber<T0>
where T1: unmanaged, INumber<T1>
where T2: unmanaged, INumber<T2>
where T3: unmanaged, INumber<T3>
where T4: unmanaged, INumber<T4>
{
return CreateTaskInternal(
left, right, result, executionHandle, 5, &InitializePacket<T0, T1, T2, T3, T4>, scaler0, scaler1,
scaler2, scaler3, scaler4);
}
/// <summary>
/// Create a task with 6 scalar parameters
/// </summary>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static MatrixTaskHandle CreateTask<T0, T1, T2, T3, T4, T5>(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
T0 scaler0,
T1 scaler1,
T2 scaler2,
T3 scaler3,
T4 scaler4,
T5 scaler5)
where T0: unmanaged, INumber<T0>
where T1: unmanaged, INumber<T1>
where T2: unmanaged, INumber<T2>
where T3: unmanaged, INumber<T3>
where T4: unmanaged, INumber<T4>
where T5: unmanaged, INumber<T5>
{
return CreateTaskInternal(
left, right, result, executionHandle, 6, &InitializePacket<T0, T1, T2, T3, T4, T5>, scaler0, scaler1,
scaler2, scaler3, scaler4, scaler5);
}
/// <summary>
/// Create a task with 7 scalar parameters
/// </summary>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static MatrixTaskHandle CreateTask<T0, T1, T2, T3, T4, T5, T6>(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
T0 scaler0,
T1 scaler1,
T2 scaler2,
T3 scaler3,
T4 scaler4,
T5 scaler5,
T6 scaler6)
where T0: unmanaged, INumber<T0>
where T1: unmanaged, INumber<T1>
where T2: unmanaged, INumber<T2>
where T3: unmanaged, INumber<T3>
where T4: unmanaged, INumber<T4>
where T5: unmanaged, INumber<T5>
where T6: unmanaged, INumber<T6>
{
return CreateTaskInternal(
left, right, result, executionHandle, 7, &InitializePacket<T0, T1, T2, T3, T4, T5, T6>, scaler0,
scaler1, scaler2, scaler3, scaler4, scaler5, scaler6);
}
/// <summary>
/// Create a task with 8 scalar parameters (maximum)
/// </summary>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static MatrixTaskHandle CreateTask<T0, T1, T2, T3, T4, T5, T6, T7>(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
T0 scaler0,
T1 scaler1,
T2 scaler2,
T3 scaler3,
T4 scaler4,
T5 scaler5,
T6 scaler6,
T7 scaler7)
where T0: unmanaged, INumber<T0>
where T1: unmanaged, INumber<T1>
where T2: unmanaged, INumber<T2>
where T3: unmanaged, INumber<T3>
where T4: unmanaged, INumber<T4>
where T5: unmanaged, INumber<T5>
where T6: unmanaged, INumber<T6>
where T7: unmanaged, INumber<T7>
{
return CreateTaskInternal(
left, right, result, executionHandle, 8, &InitializePacket<T0, T1, T2, T3, T4, T5, T6, T7>, scaler0,
scaler1, scaler2, scaler3, scaler4, scaler5, scaler6, scaler7);
}
/// <summary>
/// Get allocation statistics for monitoring pool usage
/// </summary>
public static (long TaskCount, long PacketCount, long TotalMemory) GetStatistics()
{
// Note: This would need to be implemented in SlabAllocator to get actual statistics
// For now, return placeholder values
return (0, 0, 0);
}
/// <summary>
/// Clean up all allocators (call on application shutdown)
/// </summary>
public static void Shutdown()
{
TaskAllocator?.Dispose();
ScalarPacketAllocator?.Dispose();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static void AllocateTaskAndPacket(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
int scalerCount,
out MatrixParallelTask* task,
out ScalarDataPacket* packet)
{
if (scalerCount is < 1 or > ScalarDataPacket.MaxScalarCount) {
throw new ArgumentOutOfRangeException(nameof(scalerCount));
}
// Allocate task
nuint taskPtr = TaskAllocator.Allocate();
task = (MatrixParallelTask*)taskPtr;
// Allocate scalar packet
nuint packetPtr = ScalarPacketAllocator.Allocate();
packet = (ScalarDataPacket*)packetPtr;
// Initialize task
task->left = left;
task->right = right;
task->result = result;
task->executionHandle = executionHandle;
task->scalerCount = scalerCount;
// Clear packet memory
Unsafe.InitBlock(packet, 0, ScalarDataPacket.PacketSize);
}
/// <summary>
/// Internal helper for creating tasks with scalar packets
/// </summary>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static MatrixTaskHandle CreateTaskInternal<T0>(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
int scalerCount,
delegate*<ScalarDataPacket*, T0, void> packetInitializer,
T0 param0)
where T0: unmanaged
{
AllocateTaskAndPacket(
left, right, result, executionHandle, scalerCount, out MatrixParallelTask* task,
out ScalarDataPacket* packet);
packetInitializer(packet, param0);
task->scalerDataPacket = (nint)packet;
return new MatrixTaskHandle(task, packet, true);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static MatrixTaskHandle CreateTaskInternal<T0, T1>(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
int scalerCount,
delegate*<ScalarDataPacket*, T0, T1, void> packetInitializer,
T0 param0,
T1 param1)
where T0: unmanaged
where T1: unmanaged
{
AllocateTaskAndPacket(
left, right, result, executionHandle, scalerCount, out MatrixParallelTask* task,
out ScalarDataPacket* packet);
packetInitializer(packet, param0, param1);
task->scalerDataPacket = (nint)packet;
return new MatrixTaskHandle(task, packet, true);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static MatrixTaskHandle CreateTaskInternal<T0, T1, T2>(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
int scalerCount,
delegate*<ScalarDataPacket*, T0, T1, T2, void> packetInitializer,
T0 param0,
T1 param1,
T2 param2)
where T0: unmanaged
where T1: unmanaged
where T2: unmanaged
{
AllocateTaskAndPacket(
left, right, result, executionHandle, scalerCount, out MatrixParallelTask* task,
out ScalarDataPacket* packet);
packetInitializer(packet, param0, param1, param2);
task->scalerDataPacket = (nint)packet;
return new MatrixTaskHandle(task, packet, true);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static MatrixTaskHandle CreateTaskInternal<T0, T1, T2, T3>(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
int scalerCount,
delegate*<ScalarDataPacket*, T0, T1, T2, T3, void> packetInitializer,
T0 param0,
T1 param1,
T2 param2,
T3 param3)
where T0: unmanaged
where T1: unmanaged
where T2: unmanaged
where T3: unmanaged
{
AllocateTaskAndPacket(
left, right, result, executionHandle, scalerCount, out MatrixParallelTask* task,
out ScalarDataPacket* packet);
packetInitializer(packet, param0, param1, param2, param3);
task->scalerDataPacket = (nint)packet;
return new MatrixTaskHandle(task, packet, true);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static MatrixTaskHandle CreateTaskInternal<T0, T1, T2, T3, T4>(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
int scalerCount,
delegate*<ScalarDataPacket*, T0, T1, T2, T3, T4, void> packetInitializer,
T0 param0,
T1 param1,
T2 param2,
T3 param3,
T4 param4)
where T0: unmanaged
where T1: unmanaged
where T2: unmanaged
where T3: unmanaged
where T4: unmanaged
{
AllocateTaskAndPacket(
left, right, result, executionHandle, scalerCount, out MatrixParallelTask* task,
out ScalarDataPacket* packet);
packetInitializer(packet, param0, param1, param2, param3, param4);
task->scalerDataPacket = (nint)packet;
return new MatrixTaskHandle(task, packet, true);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static MatrixTaskHandle CreateTaskInternal<T0, T1, T2, T3, T4, T5>(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
int scalerCount,
delegate*<ScalarDataPacket*, T0, T1, T2, T3, T4, T5, void> packetInitializer,
T0 param0,
T1 param1,
T2 param2,
T3 param3,
T4 param4,
T5 param5)
where T0: unmanaged
where T1: unmanaged
where T2: unmanaged
where T3: unmanaged
where T4: unmanaged
where T5: unmanaged
{
AllocateTaskAndPacket(
left, right, result, executionHandle, scalerCount, out MatrixParallelTask* task,
out ScalarDataPacket* packet);
packetInitializer(packet, param0, param1, param2, param3, param4, param5);
task->scalerDataPacket = (nint)packet;
return new MatrixTaskHandle(task, packet, true);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static MatrixTaskHandle CreateTaskInternal<T0, T1, T2, T3, T4, T5, T6>(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
int scalerCount,
delegate*<ScalarDataPacket*, T0, T1, T2, T3, T4, T5, T6, void> packetInitializer,
T0 param0,
T1 param1,
T2 param2,
T3 param3,
T4 param4,
T5 param5,
T6 param6)
where T0: unmanaged
where T1: unmanaged
where T2: unmanaged
where T3: unmanaged
where T4: unmanaged
where T5: unmanaged
where T6: unmanaged
{
AllocateTaskAndPacket(
left, right, result, executionHandle, scalerCount, out MatrixParallelTask* task,
out ScalarDataPacket* packet);
packetInitializer(packet, param0, param1, param2, param3, param4, param5, param6);
task->scalerDataPacket = (nint)packet;
return new MatrixTaskHandle(task, packet, true);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static MatrixTaskHandle CreateTaskInternal<T0, T1, T2, T3, T4, T5, T6, T7>(
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* result,
IntPtr executionHandle,
int scalerCount,
delegate*<ScalarDataPacket*, T0, T1, T2, T3, T4, T5, T6, T7, void> packetInitializer,
T0 param0,
T1 param1,
T2 param2,
T3 param3,
T4 param4,
T5 param5,
T6 param6,
T7 param7)
where T0: unmanaged
where T1: unmanaged
where T2: unmanaged
where T3: unmanaged
where T4: unmanaged
where T5: unmanaged
where T6: unmanaged
where T7: unmanaged
{
AllocateTaskAndPacket(
left, right, result, executionHandle, scalerCount, out MatrixParallelTask* task,
out ScalarDataPacket* packet);
packetInitializer(packet, param0, param1, param2, param3, param4, param5, param6, param7);
task->scalerDataPacket = (nint)packet;
return new MatrixTaskHandle(task, packet, true);
}
// Static packet initializer functions
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static void InitializePacket<T0>(ScalarDataPacket* packet, T0 s0) where T0: unmanaged, INumber<T0>
{
packet->GetScalerRef<T0>(0) = s0;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static void InitializePacket<T0, T1>(ScalarDataPacket* packet, T0 s0, T1 s1)
where T0: unmanaged, INumber<T0>
where T1: unmanaged, INumber<T1>
{
packet->GetScalerRef<T0>(0) = s0;
packet->GetScalerRef<T1>(1) = s1;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static void InitializePacket<T0, T1, T2>(ScalarDataPacket* packet, T0 s0, T1 s1, T2 s2)
where T0: unmanaged, INumber<T0>
where T1: unmanaged, INumber<T1>
where T2: unmanaged, INumber<T2>
{
packet->GetScalerRef<T0>(0) = s0;
packet->GetScalerRef<T1>(1) = s1;
packet->GetScalerRef<T2>(2) = s2;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static void InitializePacket<T0, T1, T2, T3>(ScalarDataPacket* packet, T0 s0, T1 s1, T2 s2, T3 s3)
where T0: unmanaged, INumber<T0>
where T1: unmanaged, INumber<T1>
where T2: unmanaged, INumber<T2>
where T3: unmanaged, INumber<T3>
{
packet->GetScalerRef<T0>(0) = s0;
packet->GetScalerRef<T1>(1) = s1;
packet->GetScalerRef<T2>(2) = s2;
packet->GetScalerRef<T3>(3) = s3;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static void InitializePacket<T0, T1, T2, T3, T4>(
ScalarDataPacket* packet,
T0 s0,
T1 s1,
T2 s2,
T3 s3,
T4 s4)
where T0: unmanaged, INumber<T0>
where T1: unmanaged, INumber<T1>
where T2: unmanaged, INumber<T2>
where T3: unmanaged, INumber<T3>
where T4: unmanaged, INumber<T4>
{
packet->GetScalerRef<T0>(0) = s0;
packet->GetScalerRef<T1>(1) = s1;
packet->GetScalerRef<T2>(2) = s2;
packet->GetScalerRef<T3>(3) = s3;
packet->GetScalerRef<T4>(4) = s4;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static void InitializePacket<T0, T1, T2, T3, T4, T5>(
ScalarDataPacket* packet,
T0 s0,
T1 s1,
T2 s2,
T3 s3,
T4 s4,
T5 s5)
where T0: unmanaged, INumber<T0>
where T1: unmanaged, INumber<T1>
where T2: unmanaged, INumber<T2>
where T3: unmanaged, INumber<T3>
where T4: unmanaged, INumber<T4>
where T5: unmanaged, INumber<T5>
{
packet->GetScalerRef<T0>(0) = s0;
packet->GetScalerRef<T1>(1) = s1;
packet->GetScalerRef<T2>(2) = s2;
packet->GetScalerRef<T3>(3) = s3;
packet->GetScalerRef<T4>(4) = s4;
packet->GetScalerRef<T5>(5) = s5;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static void InitializePacket<T0, T1, T2, T3, T4, T5, T6>(
ScalarDataPacket* packet,
T0 s0,
T1 s1,
T2 s2,
T3 s3,
T4 s4,
T5 s5,
T6 s6)
where T0: unmanaged, INumber<T0>
where T1: unmanaged, INumber<T1>
where T2: unmanaged, INumber<T2>
where T3: unmanaged, INumber<T3>
where T4: unmanaged, INumber<T4>
where T5: unmanaged, INumber<T5>
where T6: unmanaged, INumber<T6>
{
packet->GetScalerRef<T0>(0) = s0;
packet->GetScalerRef<T1>(1) = s1;
packet->GetScalerRef<T2>(2) = s2;
packet->GetScalerRef<T3>(3) = s3;
packet->GetScalerRef<T4>(4) = s4;
packet->GetScalerRef<T5>(5) = s5;
packet->GetScalerRef<T6>(6) = s6;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static void InitializePacket<T0, T1, T2, T3, T4, T5, T6, T7>(
ScalarDataPacket* packet,
T0 s0,
T1 s1,
T2 s2,
T3 s3,
T4 s4,
T5 s5,
T6 s6,
T7 s7)
where T0: unmanaged, INumber<T0>
where T1: unmanaged, INumber<T1>
where T2: unmanaged, INumber<T2>
where T3: unmanaged, INumber<T3>
where T4: unmanaged, INumber<T4>
where T5: unmanaged, INumber<T5>
where T6: unmanaged, INumber<T6>
where T7: unmanaged, INumber<T7>
{
packet->GetScalerRef<T0>(0) = s0;
packet->GetScalerRef<T1>(1) = s1;
packet->GetScalerRef<T2>(2) = s2;
packet->GetScalerRef<T3>(3) = s3;
packet->GetScalerRef<T4>(4) = s4;
packet->GetScalerRef<T5>(5) = s5;
packet->GetScalerRef<T6>(6) = s6;
packet->GetScalerRef<T7>(7) = s7;
}
/// <summary>
/// Task handle containing both task and scalar packet pointers
/// </summary>
public readonly struct MatrixTaskHandle : IDisposable
{
internal readonly ScalarDataPacket* ScalarPacketPtr;
internal readonly MatrixParallelTask* TaskPtr;
public readonly bool HasScalarPacket;
internal MatrixTaskHandle(
MatrixParallelTask* taskPtr,
ScalarDataPacket* scalarPacketPtr,
bool hasScalarPacket)
{
TaskPtr = taskPtr;
ScalarPacketPtr = scalarPacketPtr;
HasScalarPacket = hasScalarPacket;
}
public bool IsValid => TaskPtr != null;
/// <summary>
/// Release the allocated resources back to pools
/// </summary>
public void Dispose()
{
if (IsValid)
{
TaskAllocator.Free((nuint)TaskPtr);
if (HasScalarPacket && ScalarPacketPtr != null) {
ScalarPacketAllocator.Free((nuint)ScalarPacketPtr);
}
}
}
/// <summary>
/// Execute the task
/// </summary>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public readonly void Execute()
{
if (!IsValid) {
throw new InvalidOperationException("Invalid task handle");
}
TaskPtr->Execute();
}
/// <summary>
/// Execute the task and automatically dispose the handle
/// </summary>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public void ExecuteAndDispose()
{
try
{
Execute();
}
finally
{
Dispose();
}
}
}
}
}
StgSharp.Common/Mathematics/Linear/MatrixParallel.TaskWrapFactory.cs
+77
View File
@@ -0,0 +1,77 @@
//-----------------------------------------------------------------------
// -----------------------------------------------------------------------
// file="MatrixParallel.TaskWrapFactory.cs"
// Project: StgSharp
// AuthorGroup: Nitload Space
// Copyright (c) Nitload Space. All rights reserved.
//
// Permission is hereby granted, free of charge, to any person
// obtaining a copy of this software and associated documentation
// files (the “Software”), to deal in the Software without restriction,
// including without limitation the rights to use, copy, modify, merge,
// publish, distribute, sublicense, and/or sell copies of the Software,
// and to permit persons to whom the Software is furnished to do so,
// subject to the following conditions:
//
// The above copyright notice and
// this permission notice shall be included in all copies
// or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED “AS IS”,
// WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED,
// INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
// IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM,
// DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
// ARISING FROM, OUT OF OR IN CONNECTION WITH
// THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
//
// -----------------------------------------------------------------------
// -----------------------------------------------------------------------
using StgSharp.HighPerformance.Memory;
using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Threading.Tasks;
namespace StgSharp.Mathematics
{
internal static unsafe class MatrixParallelFactory
{
private static readonly SlabAllocator<ScalarDataPacket> _scalarAllocator = SlabAllocator<ScalarDataPacket>.Create(64, SlabBufferLayout.Chunked, true);
private static readonly SlabAllocator<MatrixParallelTaskWrap> _wrapAllocator = SlabAllocator<MatrixParallelTaskWrap>.Create(64, SlabBufferLayout.Chunked, true);
public static ScalarDataPacket* CreateScalarPacket()
{
ScalarDataPacket* p = (ScalarDataPacket*)_scalarAllocator.Allocate();
return p;
}
public static MatrixParallelTaskWrap* FromTaskWrap(MatrixParallelTaskWrap* source, int offset)
{
MatrixParallelTaskWrap* p = (MatrixParallelTaskWrap*)_wrapAllocator.Allocate();
*p = *source;
p->left += offset * p->leftStride;
p->right += offset * p->rightStride;
p->result += offset * p->resultStride;
return p;
}
public static void Release(ScalarDataPacket* source)
{
_scalarAllocator.Free((nuint)source);
}
public static void Release(MatrixParallelTaskWrap* source)
{
if (source->scalerPacket != null) {
_scalarAllocator.Free((nuint)source->scalerPacket);
}
_wrapAllocator.Free((nuint)source);
}
}
}
StgSharp.Common/Mathematics/Linear/MatrixParallel.cs
+12 -16
View File
@@ -45,23 +45,24 @@ namespace StgSharp.Mathematics
internal static class MatrixParallel internal static class MatrixParallel
{ {
private static Thread[] _threads; private static readonly Channel<IntPtr> _taskChannel = Channel.CreateUnbounded<IntPtr>();
private static Channel<IntPtr> _taskChannel = Channel.CreateUnbounded<IntPtr>();
private static int _paraState; private static int _paraState;
private static IReadOnlyList<Thread> _threads;
public static void Init() public static void Init()
{ {
int count = Environment.ProcessorCount - 4; int count = Environment.ProcessorCount - 4;
count = count < 0 ? 1 : count; count = count < 0 ? 1 : count;
_threads = new Thread[count]; List<Thread> _t = new(count);
Interlocked.Exchange(ref _paraState, 1); _threads = _t;
for (int i = 0; i < count; i++) for (int i = 0; i < count; i++)
{ {
_threads[i] = new(MatrixParallelCycle) _t.Add(new(MatrixParallelCycle)
{ {
IsBackground = true IsBackground = true
}; });
_threads[i].Start(); _t[i].Start();
} }
} }
@@ -75,13 +76,8 @@ namespace StgSharp.Mathematics
private static unsafe void MatrixParallelCycle() private static unsafe void MatrixParallelCycle()
{ {
Thread t_cur = Thread.CurrentThread; Thread t_cur = Thread.CurrentThread;
while (_taskChannel.Reader.TryRead(out IntPtr ptr)) int index = _threads.ToList().IndexOf(t_cur);
{ while (_taskChannel.Reader.TryRead(out IntPtr ptr)) { }
MatrixParallelTask* taskHandle = (MatrixParallelTask*)ptr;
// Execute the task directly
taskHandle->Execute();
}
} }
} }
StgSharp.Common/Mathematics/Scalar/ACos.cs
+28 -28
View File
@@ -1,33 +1,33 @@
//----------------------------------------------------------------------- //-----------------------------------------------------------------------
//----------------------------------------------------------------------- // -----------------------------------------------------------------------
// file="ACos.cs" // file="ACos.cs"
// Project: StgSharp // Project: StgSharp
// AuthorGroup: Nitload Space // AuthorGroup: Nitload Space
// Copyright (c) Nitload Space. All rights reserved. // Copyright (c) Nitload Space. All rights reserved.
// //
// Permission is hereby granted, free of charge, to any person // Permission is hereby granted, free of charge, to any person
// obtaining a copy of this software and associated documentation // obtaining a copy of this software and associated documentation
// files (the “Software”), to deal in the Software without restriction, // files (the “Software”), to deal in the Software without restriction,
// including without limitation the rights to use, copy, modify, merge, // including without limitation the rights to use, copy, modify, merge,
// publish, distribute, sublicense, and/or sell copies of the Software, // publish, distribute, sublicense, and/or sell copies of the Software,
// and to permit persons to whom the Software is furnished to do so, // and to permit persons to whom the Software is furnished to do so,
// subject to the following conditions: // subject to the following conditions:
// //
// The above copyright notice and // The above copyright notice and
// this permission notice shall be included in all copies // this permission notice shall be included in all copies
// or substantial portions of the Software. // or substantial portions of the Software.
// //
// THE SOFTWARE IS PROVIDED “AS IS”, // THE SOFTWARE IS PROVIDED “AS IS”,
// WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, // WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED,
// INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, // INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. // FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
// IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, // IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM,
// DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, // DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
// ARISING FROM, OUT OF OR IN CONNECTION WITH // ARISING FROM, OUT OF OR IN CONNECTION WITH
// THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. // THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
// //
//----------------------------------------------------------------------- // -----------------------------------------------------------------------
//----------------------------------------------------------------------- // -----------------------------------------------------------------------
using System; using System;
using System.Runtime.CompilerServices; using System.Runtime.CompilerServices;
@@ -36,10 +36,10 @@ namespace StgSharp.Mathematics
public static unsafe partial class Scaler public static unsafe partial class Scaler
{ {
[MethodImpl( MethodImplOptions.AggressiveInlining )] [MethodImpl(MethodImplOptions.AggressiveInlining)]
public static unsafe float ACos( float x ) public static unsafe float ACos(float x)
{ {
return MathF.Acos( x ); return MathF.Acos(x);
/* /*
int sign = 1; int sign = 1;